When university students set out to learn English as an additional language, it’s not just about internalizing a new list of words: they also have to learn common word combinations.
In order to sound fluent and natural, they’ll need to say “strong wind” but not “muscular wind,” or “reach a conclusion” but not “pull a conclusion” - a task that’s particularly challenging to learners who are stepping into terminology-heavy academic fields from economics to engineering.
So what’s the best way to learn those invaluable combinations, also known as collocations? According to a new paper from the UBC Sauder School of Business, data-driven learning, with a little boost from AI, is still the most reliable option.
According UBC Sauder lecturer Dr. Déogratias (Deo) Nizonkiza, the author of the study, the development of corpora — large, structured collections of texts and other language data — has been foundational for DDL approaches. It has happened in three major stages.
In 1961, a groundbreaking collection of texts called the Brown Corpus was developed at Brown University, and comprised 500 text samples from media, religion, fiction, science and other sources, each one containing approximately 2,000 words. At roughly one million words, it represented the first large-scale collection of real-world American English, and it allowed researchers to effectively catalogue and analyze the language and its usage.
In the 1980s, as technology advanced and computing became more widely available, the concept of data-driven learning, or DDL, was introduced. Instead of memorizing phrases from textbooks, learners could access increasingly massive databases of texts — with word counts eventually exceeding 100 million — to discover everyday language patterns.
In 2008 came the Corpus of Contemporary American English, or COCA — a one billion-word database that comprises nearly 500,000 texts from 1990 to 2019 and allows students not only to search for specific words, but to look for examples in different contexts and academic disciplines.
“That really significantly changed how we do things, specifically in the area of English for academic or professional purposes,” explains Dr. Nizonkiza. “But the searches took so much time, and people would get discouraged.”
In addition to the sluggish searches, another stumbling block was that many language instructors didn’t even know the corpora existed, let alone how to use them, and those who did often treated DDL as a secondary activity — not a core teaching tool.
Now with the advent of generative AI, or GenAI, says Dr. Nizonkiza, the process of searching for collocations has become far faster and more familiar — but that doesn’t mean educators should toss out the more traditional DDL approach.
While GenAI tools such as ChatGPT can replicate some of the functions of the existing tools — it can identify and generate examples of common word combinations, for example — it’s often unclear where the examples come from, or how accurate they are.
For the French-language study, whose title translates to At the Intersection of Corpora and Artificial Intelligence: A Critical Review of DDL Approaches to Teaching Collocations in Higher Education , Dr. Nizonkiza examined empirical evidence, including meta-analyses and systematic reviews, on the effectiveness of DDL.
The literature shows that a hybrid approach that’s rooted in corpus-based teaching, and gets an added boost from GenAI, is the most effective.
"We have to be very careful, because GenAI tools aren’t as precise as those traditional tools,” explains Dr. Nizonkiza. "So instead of relying perhaps on GenAI, we need to make sure that there is a combination of GenAI and the traditional corpora."
The paper also provides a possible roadmap for that hybrid approach. For example, an English instructor could select a particular vocabulary — from the Economics Academic Word List, say — then identify common collocations like “control volatility”, “market volatility,” etc… From there, they can ask corpus-informed GenAI tools to generate examples of those collocations, which can be used to design structured exercises where students first fill in the blanks (e.g., “Portfolio diversification is a key strategy used by fund managers to ______ volatility in emerging markets”), then use AI to verify their answers.
So rather than relying on popular AI platforms like ChatGPT or Copilot to generate examples, students complete the exercises — and only then use AI to verify and refine their conclusions.
That process helps students to raise awareness of collocations, improve accuracy, and spur critical interactions with GenAI tools, says Dr. Nizonkiza.
“It’s when they start university that they have exposure to these new words they haven't necessarily encountered in their everyday lives,” says Dr. Nizonkiza. “That's where having these tools, and being aware of what these tools can help them do, becomes essential.”
Nouvelles perspectives en sciences sociales
L’enseignement des collocations en anglais au niveau tertiaire : revue critique des approches DDL entre corpus et intelligence artificielle | At the Intersection of Corpora and Artificial Intelligence: A Critical Review of DDL Approaches to Teaching Collocations in Higher Education
30-Jun-2026